A new machine learning approach shows promise to help astronomers infer the internal structure of star nurseries from telescopic observations and glean new insights into star formation
bring light to the darkness
In galaxies, stars are born in huge clouds of interstellar gas and dust called molecular clouds. Through turbulence and gravitational instability, these molecular clouds develop complex substructures within which star-forming sites are located. The densest and smallest of these structures are star-forming nuclei, the direct ancestors of individual stars or entire star clusters. Inferring the structure of these clouds and cores is an important aspect of understanding star formation.
However, observing these regions is a difficult task because the dust within these clouds absorbs visible light, making the interior opaque to conventional observations. Infrared and submillimeter telescopes can be used to measure the thermal radiation emitted by the dust itself, but these observations only return a two-dimensional projection onto the sky. Therefore, it remains difficult to recover the true three-dimensional distribution of material within these clouds.
Our research team is developing a new approach that shows great promise in revealing the hidden structure of cloud cores reconstructed from observed dust emissions.
Combining simulation and machine learning
Our research combines astrophysical simulations and modern machine learning techniques to interpret dust emission observations. Specifically, we developed a framework that employs reversible neural networks. (1,2)a class of deep learning models tailored towards solving inverse problems. These are tasks that infer properties of a system that cannot be directly measured from secondary observations.
To train the system, we employed an advanced numerical simulation of star-forming clouds called Cloud Factory. (3,4) These simulations consistently track the formation of molecular clouds on the galactic scale and take into account the effects of galactic-scale forces, gas chemistry and cooling, and supernova feedback.
We then selected core-like structures from these simulations and processed this data with the radiative transfer code POLARIS. (5,6) Generate synthetic dust observations at different wavelengths, mimicking the types of data acquired by infrared and submillimeter observatories. Our neural network was then trained to learn the relationship between these synthetic dust emission observations and the underlying three-dimensional structure of the cloud core, specifically to infer both dust density and dust temperature distribution.
Accuracy and robustness
Tests on a never-before-seen example of a synthetic cloud core reveal that our new approach can reproduce internal dust density and dust temperature structure with high accuracy (Fig. 1). Even if we limit the number of wavelengths of the input dust emission observations to a range comparable to that available with existing telescopes for some of the neighboring star-forming nuclei, our method is able to highly constrain the dust structure. Our research team is currently preparing application to real observational data, but additional optimization is required to carefully close the gap between synthetic dust emission observations and real telescope data.
Another central aspect of our reversible neural network approach is the ability to estimate uncertainty. Rather than returning only one prediction of dust density and temperature for every pixel in a three-dimensional structure, our method estimates the entire probability distribution of these quantities. This provides additional insight into any variations in the dust distribution that are consistent with the input data. This is important because the available observations often cannot uniquely determine the three-dimensional structure of the dust.
Astronomy in the era of big data
Our research builds on the growing need in astronomy for highly efficient and reliable analytical methods as modern astronomical observations become increasingly detailed. Spaceborne observatories such as the Herschel Space Observatory and the James Webb Space Telescope provide vast amounts of high-quality multiwavelength dust emission observations, and future missions will no doubt generate even larger datasets.
New evaluation strategies based on machine learning techniques such as ours offer a promising means to make full use of this rich observational data. Compared to traditional analysis pipelines that can be prohibitively time-consuming, machine learning tools provide statistically reliable and highly efficient inference methods on large datasets after training.
Machine learning as a link between theory and observation
A central pillar of astronomical research is the comparison of theoretical models with actual observations. Simulation-based inference methods on observational data, such as our machine learning approach, serve to efficiently connect the realms of theory and observation. On the one hand, it allows us to interpret vast amounts of observational data through the lens of theoretical understanding. But on the other hand, it can also highlight areas where our theoretical models may not be sophisticated enough to accurately describe reality.
For example, our study found that the synthetic dust emission observations initially produced were too simplistic to accurately match actual observations of star-forming nuclei. (7) Improving these modeling efforts is therefore a central part of our ongoing research into reconstructing dust distributions in star-forming nuclei.
References
- Ardizzone et al. (2018): “Analysis of inverse problems using reversible neural networks”, International Conference on Learning Representations 2019.
- Ardizzone et al. (2019): “Guided Image Generation with Conditional Reversible Neural Networks”, arXiv:1907.02392
- Smith et al. (2020): “Cloud Factory I: Producing resolved filamentary molecular clouds from galactic-scale forces”, Monthly Notices of the Royal Astronomical Society, 492, 1594.
- Izquierdo et al. (2021): “Cloud Factory II: Heavy turbulent kinematics of resolved molecular clouds in the galactic potential”, Monthly Notices of the Royal Astronomical Society, 500, 5268.
- Reisl et al. (2016): “Radiative transfer with POLARIS. I. Analysis of magnetic fields by synthetic dust continuum polarization measurements, Astronomy & Astrophysics, 593, A87.
- Reisl et al. (2019): “Radiative Transfer with POLARIS” II. “Modeling of Synthetic Galaxy Synchrotron Observations”, The Astrophysical Journal, 885, 15.
- Kusol et al. (2024): “A deep learning approach to 3D reconstruction of dust density and temperature in star-forming regions”, Astronomy and Astrophysics, 683, A246.
